如何在以下代码中改进读写的位置？

Question

我正在处理以下图片卷积代码：

typedef struct fmatrix{
    int rows;
    int cols;
    float** array;
} fmatrix;

typedef struct image{
    unsigned char* data;
    int w;
    int h;
    int c;
} image;

typedef struct kernel{
    fmatrix* psf;
    int divisor;
} kernel;

void convolve_sq(image* src, image* dst, kernel* psf, int pixel){
    int size = psf->psf->rows * psf->psf->cols;
    float tmp[size];
    int n, m; //for psf
    int x, y, x0, y0, cur; //for image

    y0 = pixel / (src->w * src->c);
    x0 = (pixel / src->c) % src->w;

    for (n = 0; n < psf->psf->rows; ++n){
        for (m = 0; m < psf->psf->cols; ++m){
            y = n - (psf->psf->rows / 2);
            x = m - (psf->psf->cols / 2);
            if ((y + y0) < 0 || (y + y0) >= src->h || (x + x0) < 0 || (x + x0) >= src->w){
                tmp[n*psf->psf->rows+m] = 255 * psf->psf->array[n][m];
            }
            else{
                cur = (pixel + y * src->w * src->c + x * src->c);
                tmp[n*psf->psf->rows+m] = src->data[cur] * psf->psf->array[n][m]; //misses on read
            }
        }
    }
    m = 0;
    for (n = 0; n < size; ++n){
        m += (int) tmp[n];
    }
    m /= psf->divisor;
    if (m < 0) m = 0;
    if (m > 255) m = 255;
    dst->data[pixel] = m; //misses on write
}

void convolve_image(image* src, image* dst, kernel* psf){
    int i, j, k;
    for (i = 0; i < src->h; ++i){
        for (j = 0; j < src->w; ++j){
            for (k = 0; k < src->c; ++k){
                convolve_sq(src, dst, psf, (i * src->w * src->c + j * src->c + k) );
            }
        }
    }
}

运行cachegrind，我已经确定了两个存在大量缓存未命中的地方，我已在上面的代码中注释了这些地方。对于标记为“读取未命中”的行，有107,205个D1mr和97,201个DLmr。对于标记为“写入未命中”的行，有107,201个D1mw和DLmw。这些行分别直接读取和写入图像。

如何在避免缓存未命中方面提高此代码的效率？